What Maple-Preview is
Maple-Preview is a large language model from DeepGrove specialised for reasoning — deriving an answer from an input. It uses ternary weights, restricting weight values to three, and was designed from the start to run on the machine in front of you.
The configuration is written as 20B-A1B: 24 layers with 256 experts, of which only 8 are active for a given pass. What that notation conveys is a model that is large as a whole but only partly active at any moment, which is the basis of both its speed and its low memory use.
The checkpoint — the distributed file containing the trained model itself — is 5.31GB, and the context it handles is 131,072 tokens. DeepGrove says it can solve problems at International Mathematical Olympiad level, claiming state-of-the-art reasoning for its weight class and competitiveness with larger models.
Today we introduce Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM. ／ Maple-Preview has SOTA reasoning for its weight class and is even competitive with larger models. ／ 218 tok/s M4 Mac mini ／ 5.31 GB Checkpoint ／ 131,072 Token context ／ Maple-Preview is a 20B-A1B reasoning model designed from the start for efficient on-device inference. ／ It utilizes a 24-layer, 256-expert (8 active) configuration with 3:1 SWA-512:GA attention. — From the passages on the model's position and release form, the claimed speed, size and context length, and the layer and expert configuration
How to read the speed figure
The speed is what draws attention. DeepGrove cites over 200 tokens per second on a Mac mini with an M4, published as 218 tokens/sec, and describes this as 5–16x faster than efficiency-focused models such as Gemma 4, Qwen3.5 and gpt-oss.
There is a caveat attached, though. The Transformers implementation included with the model card depends on Triton and FlashAttention and is intended for a CUDA environment — the runtime for NVIDIA GPUs. The Apple Silicon figure was not produced with that implementation but with a separate on-device runtime.
So this does not read as "run the model card's code on a Mac and you get 218 tokens/sec". Reproducing that speed requires a different runtime that has not been published. Missing this point undermines the premise of any evaluation.
It solves IMO-level problems and runs at 200+ tokens/sec on a Mac mini M4, 5–16× faster than efficient models like Gemma 4, Qwen3.5, and gpt-oss. ／ The included Transformers implementation depends on Triton and FlashAttention and is intended for a compatible CUDA environment. ／ The reported Apple Silicon result uses a separate on-device runtime. — From the claimed speed and the models compared against, the dependencies and intended environment of the included implementation, and the note on what the Apple Silicon result was measured on
The weakness DeepGrove admits to
On evaluation, DeepGrove says the model sets a new point on the Pareto frontier for both memory-to-performance and speed-to-performance. In the same breath it writes that this preview is focused on raw reasoning and may underperform on agentic benchmarks.
The reason is spelled out. Post-training for agentic tasks — the additional training layered on after the base training to fit a purpose — was minimal, and general reinforcement learning, which rewards good outputs to shape behaviour, was only small-scale. Training is to continue before the full release.
Delegating reasoning and delegating tool use are different jobs. If you are picking a small model to run locally, they are worth assessing separately.
On benchmarks, Maple-Preview sets a new point on the Pareto frontier for both memory-to-performance and speed-to-performance, demonstrating its strong reasoning capabilities. ／ However, we note that this preview is focused primarily on raw reasoning and, as such, may underperform on agentic benchmarks. ／ We intend to continue improving general performance through extended training before Maple's full release. ／ This preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning. ／ Maple-Preview is released under the [MIT License](LICENSE). — From the passages on where the evaluation results place the model, the limitation on agentic benchmarks and its cause, the plan ahead of full release, and the licence
How to read Maple-Preview
Maple-Preview is a preview that goes all in on fast, light reasoning. Handling 131,072 tokens in 5.31GB, under a licence that permits commercial use, is a strong position on its own.
Against that, the Mac speed presumes a different runtime, and DeepGrove itself calls the agentic side unfinished. If you are choosing a small model to run locally, setting it beside LFM2.5-2.6B — released around the same time and aimed squarely at agentic use — makes the difference obvious. For production use, waiting for the full release is the safer call.
